Electron Correlation Methodology - American Chemical Society

Chapter 5

Bond Breaking in Quantum Chemistry: A Comparison of Single- and Multi-Reference Methods

Downloaded by UNIV QUEENSLAND on June 4, 2013 | http://pubs.acs.org Publication Date: March 13, 2007 | doi: 10.1021/bk-2007-0958.ch005

C . David Sherrill, Antara Dutta, Micah L. Abrams, and John S. Sears Center for Computational Molecular Science and Technology, School of Chemistry and Biochemistry, Georgia Institute of Technology, Atlanta, GA 30332-0400

Standard and new quantum chemical methods are evaluated for their ability to provide accurate potential energy curves for chemical reactions which break or form bonds. Comparisons to full configuration interaction benchmark results demon strate that even high level single-reference methods such as unrestricted coupled-cluster with single, double, and perturbative triple substitutions [UCCSD(T)] can have large errors for bond breaking processes: nonparallelity errors for breaking bonds to hydrogen, which should be one of the simplest theoretical problems, are around 3-4 kcal mol . Multi-reference methods are much more reliable but are also more computationally expensive. New, minimalist config uration interaction methods for bond breaking in larger molecules are also discussed which dramatically improve on similar earlier models. -1

The Bond Breaking Problem The vast majority of quantum chemical studies focus on equilibrium prop erties. However, a detailed understanding of chemical reactions requires a description of their chemical dynamics, which in turn requires information about the change in potential energy as bonds are broken or formed. Even though modern electronic structure theory can provide near-spectroscopic accuracy for small molecular systems near their equilibrium geometries, the general description of potential energy surfaces away from equilibrium remains very much a frontier area of research. © 2007 American Chemical Society

In Electron Correlation Methodology; Wilson, A., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2007.

75

76


Nearly all of the commonly used quantum chemical methods are ultimately based upon Hartree-Fock molecular orbital theory, which describes the motion of each electron in the average field of all the other electrons. This introduces an error because in reality the motions of electrons are correlated. Short-range electron correlation, referred to as "dynamical" correlation, is well described by the hierarchy of many-body perturbation theory and coupled-cluster methods. However, long-range electron correlations, referred to as "nondynamical" correlation, are also important and become crucial for a proper description of bond breaking processes. At the dissociation limit, the energy due to nondynamical correlation can be larger than the energy due to dynamical correlation (7). The origin of this nondynamical correlation is that as bonds are broken, it is no longer true that a single electron configuration is sufficient as a zeroth-order wavefunction. It is well known that Hartree-Fock with restricted orbitals is incapable of providing qualitatively correct potential energy curves for bond breaking reactions. For the H molecule, for example, the (a ) electron configuration is appropriate at equilibrium, but at dissociation, the a and σ[ orbitals become degenerate, necessitating an equivalent treatment of the (σ^) and (σ[) configurations which is not provided by Hartree-Fock theory. Although additional configurations such as (σ[) are added to the wavefunction via standard "single-reference" correlation methods, for larger molecules the selection of additional configurations will be imbalanced because they are generated relative to the single Hartree-Fock reference only. A completely balanced treatment o f electron correlation would require, for example, all single and double excitations out of all near-degenerate electron configurations. Such an approach is an example of a multi-reference method. 2

2

g

g

2

2

2

Although multi-reference methods are capable of accurately describing bond breaking processes in principle, they are very difficult to formulate, implement, and apply. Typically, their computational cost is prohibitive for all but the smallest molecules. Additionally, they can require seemingly arbitrary choices of reference configurations or active spaces. Hence, continued development of methods for bond breaking has emphasized both more efficient multi-reference methods as well as generalized single-reference methods. Within the scope of single-reference methods, there are two obvious, complementary strategies to attack the bond breaking problem. First, one may simply include more and more electron configurations (e.g., triple, quadruple substitutions) to approach the limit in which all possible configurations are included. Unfortunately, the computational cost of this strategy increases rapidly. Second, one may abandon restricted orbitals in favor of unrestricted orbitals. Unrestricted Hartree-Fock (UHF) provides qualitatively correct dissociation curves for molecules in which a single bond is broken, even though it can be quantitatively poor. B y adding sophisticated treatments of electron



77 correlation, one might hope to obtain quantitatively good potential energy curves. However, using unrestricted orbitals introduces spin contamination, and the wavefunction is no longer an eigenfunction of the total spin operator. Any spin-dependent properties will no longer be properly predicted. One of the most important steps in the development of new methods for bond breaking reactions is the calibration of the method against reliable benchmark results. Unfortunately, these benchmarks are very hard to obtain. In the general case, there is no simple way to obtain entire potential energy curves directly from experiment; hence, theoretical benchmarks are required. This problem may be addressed by solving the electronic Schrôdinger equation exactly within the Born-Oppenheimer approximation, using the full configuration interaction (FCI) model. Because the computational cost of full CI increases factorially with the number of electrons or orbitals, it can be used only for the smallest chemical systems. Nevertheless, advances in algorithms and computer hardware have recently made it possible to obtain full CI potential energy curves for a few small molecules (2-5). In particular, Olsen and co workers have presented very enlightening benchmark curves for several electronic states of N ( 5 , 6) and full CI energies for five points along the double dissociation of H 0 (J). A number of standard theoretical methods, including many-body perturbation theory and coupled-cluster theory, were compared to the full CI results. Benchmark curves such as these are essential for calibrating new theoretical methods meant to provide accurate potential energy curves (712). 2

2

At present, there remains a scarcity of high quality full CI potential energy curves. While Olsen and co-workers have focused on the most challenging cases of double and triple bond breaking, we have examined (13,14) more typical cases which should also be less challenging for standard methods. Various single- and multi-reference methods have been compared to our benchmark full CI results. We find rather large errors in even the most advanced of the standard single-reference methods. Results for the multi-reference methods considered are much improved. Whether accurate results can be obtained for bond breaking using com putationally inexpensive methods remains an open question. One new strategy, described elsewhere in this volume, is the spin-flip approach of Krylov and co workers (12, 15, 16). This is a single-reference approach which attempts to avoid the usual pitfalls of such methods by using a high-spin (M = 1) triplet reference state to generate determinants appropriate for a ground state singlet (M = 0) potential energy curve. O f course generating a M , = 0 determinant from an M , = 1 determinant requires a spin flip, and hence the name of the method. The motivation for this approach is the supposition that the = 1 triplet state is easier to describe at the Hartree-Fock level across the whole potential energy curve. At its simplest level, only single (spin-flipping) excitations are allowed s

s


78


relative to the high-spin triplet reference; this is called the SF-CIS or SF-SCF method, and it is meant to give singlet potential energy curves with an accuracy roughly comparable to that of the Hartree-Fock approach, except that this accuracy is now roughly uniform across the entire curve. Improvements can be made by including various models of dynamical electron correlation (12, 16). One drawback of this approach is that the wavefunctions it generates are not spin eigenfiinctions. We have recently examined (17) the effect of adding the requisite determinants to obtain spin eigenfiinctions, and we find that this makes dramatic improvements.

An Assessment of Single-Reference Methods The performance of standard single-reference electronic structure methods has not been fully assessed for bond breaking reactions because full config uration interaction benchmarks are not widely available for entire potential energy curves. In recent work (13), we have obtained full CI potential energy curves for breaking single bonds. In particular, we have used the DETCI program (18) in the PSI3 (19) package to obtain benchmark full CI curves for breaking a bond to hydrogen in B H , H F , and CH4 using the basis sets aug-cc-pVQZ, 631G**, and 6-31G*, respectively. We prefer the latter Pople basis sets among standard polarized double zeta sets because they give the best predictions of molecular properties at the full CI level (20). Methods compared include Hartree-Fock theory, second-order M0ller-Plesset perturbation theory (MP2) (21); coupled-cluster with single and double substitutions (CCSD) (22); and coupled-cluster with single, double, and perturbative triple substitutions [CCSD(T)] (23). We have also examined the B 3 L Y P gradient-corrected hybrid density functional theory model (24,25). Both restricted and unrestricted orbitals have been used. Our results were qualitatively similar for the B H , H F , and C H molecules, so we will focus here on C H as a representative case. The single-reference potential energy curves are compared to full CI in Figure 1, in which one C - H bond is stretched while the other geometrical parameters are held constant. One immediately notices the well-known failure of restricted Hartree-Fock (RHF) at large bond lengths; this is caused by the presence of unphysical, high-energy ionic terms in the R H F energy which may be removed by employing unrestricted orbitals or by using an appropriate multiconfigurational selfconsistent-field (MCSCF) approach (26). Unfortunately, the failure of R H F is so severe that none of the correlated methods based upon it are able to overcome it (except, of course, full CI). The M P 2 curve diverges to negative infinity at the dissociation limit because of a near-degeneracy between the highest occupied and lowest unoccupied molecular 4

4



79

-40.45 A 0.5

1 !

ι 1.5

1 2

1 2.5

1 3

1

1 • 35

1

r—* 4

1 45

R C H (angstrom)

Figure 1. Potential energy curves for CH in a 6-31G* basis using restricted orbitals. (Data from reference 13.) 4

orbitals, and the perturbative treatment of triple substitutions in CCSD(T) also causes it to fail at large bond lengths. The density functional theory method B 3 L Y P is seen to suffer the same fate as Hartree-Fock when the orbitals are restricted: energies for large bond lengths are much too large. O f all restricted methods considered, only C C S D provides results which are qualitatively correct, but the quantitative error at large bond lengths is around 12 kcal mol" . One might hope that highly correlated methods employing an unrestricted Hartree-Fock reference might be able to overcome these difficulties, since U H F gives a qualitatively correct (if quantitatively poor) description of bond breaking. Results for U H F references are plotted in Figure 2, where the acronyms for methods have been prefixed by ' U ' to denote unrestricted orbitals. The U H F curve levels off too quickly with increasing bond length (underestimating the dissociation energy); however, the U C C S D and UCCSD(T) curves appear close to the full CI benchmark curve. Similarly, although U B 3 L Y P provides energies much lower than full CI for this basis set, the potential energy curve has a shape close to that of full CI. Since relative energies, not total energies, are the only relevant quantity for chemical reactions, we are only concerned with how well the curves parallel the full CI results. O f the correlated methods, only U M P 2 provides a curve which is qualitatively 1


80 -40.10

-40 15

-40.20

î t

-40 25

£ -40.30 U


•40 35 -40.40 15

-40450.5

2

2.5 R C H (angstrom)

4.5

3.5

3

Figure 2: Potential energy curves for CH in a 6-31G* basis using unrestricted orbitals. (Data from reference 13.) 4

120

no too 90

— · — RHF-FC1

80

-•-MP2-FC1

70

-*-CCSD(T)-FCI

-•-CCSD-FCI -W-B3LYP-FCI+70

·-

-•-UHF-FCI -CKUMP2-FCI -•-UCCSD-FCI -FCT -f-UB3LYP.FC1*70

^ *

—*-

m

t l 1 1 1

05

•

1

*

*^ ' ι

ι

ι

p

y—Β—Ϊ—ι—

ι

ι

ι

1 ι

ι

1

ι

-I

—

2

25

3

45

3.5

R C H (angstrom)

F/gwre 5: Errors m potential energies for CH in a 6-31G* basis. B3LYP and UB3LYP error curves have been shifted up by 70 kcal mot . (Data from reference 13.) 4

1


81 different from full CI. The U M P 2 energy clearly rises too rapidly in the intermediate bond breaking region, between about 1.6 and 2.2 À. A more detailed examination of the errors in the approximate single ref erence methods for CH4 is provided in Figure 3, which plots errors vs. full CI as a function of bond length. A completely flat error curve would be perfectly parallel to the full CI curve and would provide exactly the same predictions for properties such as bond length, harmonic frequency, dissociation energy, etc. Note that restricted and unrestricted curves coincide before a certain bond length which depends on the method. While all the approximate restricted methods except C C S D have very large errors near dissociation, the unrestricted methods have their largest errors in the intermediate region. This error is very large for U M P 2 , rising from around 15 kcal mol* near equilibrium to more than 30 kcal mol* in the intermediate bond breaking region. At these geometries, the U M P 2 error is larger than that for restricted MP2. Besides C C S D , the only methods with small errors are U C C S D and UCCSD(T). In order to help quantify the errors, we have computed the so-called non-parallelity error (NPE), which is the magnitude of the difference between the largest and smallest errors along the potential energy curve. A curve perfectly parallel to full CI would have a constant error across the potential energy curve, and therefore an N P E of zero. N P E values for C H , as well as B H and H F , are presented in Table 1. The NPEs for C C S D , U C C S D , and UCCSD(T) for CH are 10.3, 5.1, and 3.2 kcal mol" , respectively. None of these errors are small enough to claim that these methods provide "chemical accuracy," which is commonly taken to mean errors in relative energies of about 1-2 kcal mol" . This is perhaps surprising since one might expect breaking a bond to a hydrogen atom would be one of the easiest possible bond breaking problems. Table 1 shows that the errors for C H are fairly similar for the B H and H F molecules. UCCSD(T) provides the best results in all cases, yet the error never drops below 3 kcal mol . U C C S D is the only alternative i f one desires even a modest accuracy of ± 10 kcal mol" .


1

1

4

1

4

1

4

1

An Assessment of Multi-Reference Methods B y treating all near degenerate electron configurations on an equal footing, multi-reference methods are capable of handling the bond breaking problem, albeit at an increased cost in theoretical complexity and computer time. We have recently compared (14) several multi-configurational reference functions and multi-reference methods based upon them for bond breaking in a few simple molecules including the cases considered above: B H , HF, and C H . In this section we will discuss our general findings and examine C H in particular. 4

4


82 1

Table 1: Non-parallelity error (NPE, kcal mol' ) for BH, H F and C H molecules.


4

Method RHF MP2 CCSD CCSD(T) B3LYP UHF UMP2 UCCSD UCCSD(T) UB3LYP CASSCF CASPT2 CISD[TQ] SOCI

BH 8.1 23.3 n/a 27.1 17.9 4.7 3.1 n/a 9.4 1.8 0.3 0.3

HF 12.9 66.7 47.8 25.9 6.0 3.7 5.9 18.0 1.6 5.3 5.3

CH 10.3 63.0 27.4 17.1 5.1 3.2 11.4 6.3 0.7 1.3 0.3 4

NOTE: n/a denotes data not available, dashes denote very large or divergent errors. SOURCE: Data from Refs. 13,14.

Figure 4: Errors in potential energies for CH in a 6-31G* basis using multi-reference methods. (Data from reference 14.) 4


83 Figure 4 presents results for breaking a C - H bond in CH4 obtained with several multi-reference methods. First, we have used the complete-active-space self-consistent-field (CASSCF) method (26), which includes all possible electron configurations which can be generated in a subset of orbitals termed the "active space." We have used the space of all valence orbitals as our active space; this choice should allow C A S S C F to describe any bond breaking process at least qualitatively correctly by including all configurations which could become near-degenerate. The C A S S C F errors plotted in the figure have been shifted down by 40 kcal mol* for an easier comparison with the other methods. The error is largest around equilibrium, and flattens out at large distances. This happens because the dynamical electron correlation, not described properly by C A S S C F , is larger at short distances, where more electrons are closer together. The next method considered is CASPT2 (27), which adds a second-order perturbation theory treatment of dynamical correlation on top of the qualitatively correct treatment of nondynamical correlation in C A S S C F . The CASPT2 curve is much closer to full CI, as indicated by a much flatter error curve in Figure 4. The nonparallelity errors of C A S S C F and CASPT2 in Table 1 are 6.3 and 0.7 kcal mol" for CH4, and these are seen to be fairly typical of the molecules considered (errors for HF are larger because a valence active space for that molecule includes only a single antibonding orbital). This N P E for C A S S C F is roughly comparable to that of C C S D , but it should be pointed out that C A S S C F is more robust for these problems and is less likely to fail for more challenging cases. The CASPT2 results are superior to any of the single-reference methods considered above. To obtain an even greater accuracy, one might turn to multi-reference configuration interaction or multi-reference coupled-cluster methods. These approaches are typically less straightforward to apply than C A S S C F or CASPT2, because in addition to the active space, one may also need to choose thresholds or reference determinants. We favor multi-reference methods that choose determinants in a simple, a priori fashion, such as in the restricted active space configuration interaction (RASCI) approach of Olsen et ai (28). Here, we have looked at a particular multi-reference CI approach which can be formulated as a RASCI, which is the so-called second-order CI (SOCI). This method generates all single and double excitations from every determinant in a C A S S C F wavefunction. Unfortunately, this procedure generates a tremendous number of determinants, and few SOCI computations have been performed. For this reason, we have also considered an approximation to SOCI in which all determinants which would be considered more than quadruple excitations relative to the dominant Hartree-Fock reference determinant are discarded. This approach, introduced by Schaefer and co-workers (29-31), is designated CISD[TQ] because it includes all single and double substitutions but only some triples and quadruples. This wavefunction can also be formulated in the context of a RASCI.


1

1


84 The SOCI and CISD[TQ] curves for C H are so close to full CI that they are difficult to distinguish from fiill CI or from each other. However, they may be examined more clearly in the error curves in Figure 4, which display very flat errors for SOCI, while CISDfTQ] errors increase gradually with increasing bond length. This suggests that some of the triple, quadruple, or higher excitations present in SOCI but neglected in CISD[TQ] are becoming important at large bond lengths. This is perhaps surprising in that, for C H , there should be only two dominant configurations at large bond lengths, and the CISD[TQ] wavefiinction certainly includes these as well as all single and double excitations from each of them. On the other hand, the wavefiinction also includes more triples and quadruples relative to the configuration dominant at equilibrium than to the antibonding one which becomes degenerate at the dissociation limit, suggesting a possible minor imbalance in the method. At any rate, the absolute errors for CISD[TQ] and SOCI are much smaller than for any other methods considered here, and their nonparallelity errors for CH4 are very impressive at 1.3 and 0.3 kcal mol" , respectively. These methods should prove extremely reliable for any single bond breaking process where they can be afforded, and in principle, SOCI is capable of breaking any number of bonds simultaneously. 4


4

1

Minimalist Configuration-Interaction Approaches As the above test cases demonstrate, the most sophisticated treatments of electron correlation are the most successful at describing bond breaking processes. However, it is important to ask whether one can formulate much simpler, less computationally demanding approaches which might still describe bond breaking accurately. As discussed in the introduction, the spin-flip technique (12, 15, 16) is one attempt to adapt standard single-reference electronic structure methods in a simple way to be more appropriate for bond breaking. The simplest of these models, SF-SCF (also called SF-CIS), obtains singlet wavefimctions from a high-spin triplet reference by generating all single substitutions which flip a single spin from a to β. One drawback of this approach as originally formulated is that the generated set of determinants does not include all determinants necessary to form eigenfiinctions of S . Recently, two of us (C.D.S. and J.S.S.) implemented the ability to obtain the spin-complete analogues of SF-SCF, denoted here SC-SF-SCF, into our group's DETCI program. One advantage of our implementation is that it decouples the orbitals used from the identity of the reference determinant. Hence, it is possible to generate spin-flipped singlet determinants relative to a triplet "reference" determinant but use, for example, closed-shell singlet orbitals. This allows us to examine the effects of orbitals separately from the selection of determinants. 2


85 Table 2: Equilibrium bond lengths and dissociation energies for the F molecule with a DZP + basis set.

e

2

Method

Re

De

CCSD UCCSD SF-SCF SC-SF-SCF (singlet orbitals)' SC-SF-SCF (triplet orbitals)' SF-CIS(D)'' SF-OD'' VOO-CCD(2/ MR-CISD Expt.

1.332 1.410 1.410 1.567 1.469 1.448 1.429 1.437 1.417 1.435 1.412

10.69 2.36 0.95 0.28 1.29 1.37 1.14 1.24 1.51 1.22 1.66

C

f


7

C

e

3

Spin-flip approaches used a Σ„ reference.

^Computed at r(F-F) = 100 bohr. Reference 32. ^Reference 12. Reference 17/Reference 9.

Having spin complete wavefunctions dramatically improves the accuracy of the SF-SCF method. One of the most striking examples of this is for the very challenging F molecule, considered in Table 2. As usual, the R H F approach is incapable of properly dissociating the molecule and gives unreasonable dissociation energies. Even single-reference C C S D methods perform very badly for this difficult case (32), giving good bond lengths but very poor dissociation energies, in contrast to the reasonable performance of C C S D for breaking single bonds to hydrogen atoms discussed above. Multi-reference CISD and the V O O CCD(2) model (8) both give good results for this case (9, 32), considering the limited basis set used. Although SF-SCF improves significantly over the unphysical results of RHF, nevertheless its performance is poor, overestimating r by .15 À and predicting a dissociation energy of 0.28 eV compared to an experimental value of 1.66 eV. The spin-complete alternative, however, gives much better results: the error in the bond length is reduced by about two thirds, and the dissociation energy becomes much more reasonable around 1.3 eV. In this case, triplet orbitals are more effective than singlet orbitals, giving dissociation energies of 1.37 eV vs. 1.29 eV. Closer agreement with experiment requires a treatment of dynamical electron correlation. Since SF-CIS(D) was found to improve so dramatically over simple SF-SCF, we anticipate perturbative corrections for dynamical correlation in our spin-complete version should be even more effec2

e


86 tive. It should be noted that the most extensive spin-flip model considered for this problem, SF-OD, gives results (12) which rival M R - C I S D for this case. A perhaps more typical comparison of SF-SCF and its spin-complete alternative is presented in Figure 5, which presents potential energy curves for the H F molecule. One sees that using spin complete waveftmctions reduces the error in the SF-SCF energies by roughly 2/3 at large distances.

018

0.16


9-9 014

012

I

01

/

5 008

0.06

0.04

0.02 |-

9

Λ/'

I

///

SC-SF-SCF (stnglet orbs) SC-SF-SCF (triplet orbs) SF-SCF --: SF-CISgg - FCI ·

if 05

2.5

15

35

R (Angstrom)

Figure 5: Potential energy curves for HF in a 6-31G basis. (Data from reference 17.)

We anticipate that continued development of minimalist models, in con junction with more efficient multi-reference approaches and a wider array of benchmark curves, will result in a better understanding of the theoretical challenges of bond breaking processes and increased ability of electronic structure theory to model them.

Acknowledgments C.D.S. acknowledges a Camille and Henry Dreyfiis New Faculty Award and an N S F C A R E E R Award (Grant No. N S F 0094088). The Center for Computational Molecular Science and Technology is funded through a Shared University Research (SUR) grant from I B M and by Georgia Tech.


87

References 1. 2. 3. 4.


5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19.

20. 21. 22. 23. 24. 25. 26. 27.

Shavitt, I. in Methods of Electronic Structure Theory, Schaefer, H . F., Ed., pp 189-275. Plenum Press, New York, 1977. van Mourik, T.; van Lenthe, J. H . J. Chem. Phys., 1995, 102, 7479. Olsen, J.; Jørgensen, P.; Koch, H . ; Balková, Α.; Bartlett, R. J. J. Chem. Phys., 1996, 104, 8007. Krylov, A . I.; Sherrill, C. D.; Byrd, E. F. C.; Head-Gordon, M . J. Chem. Phys., 1998, 109, 10669. Larsen, H.; Olsen, J.; Jørgensen, P.; Christiansen, O. J. Chem. Phys., 2000, 113, 6677. Larsen, H.; Olsen, J.; Jørgensen, P.; Christiansen, O. J. Chem. Phys, 2001, 114, 10985. Laidig, W. D.; Bartlett, R. J. Chem. Phys. Lett., 1984, 104, 424. Gwaltney, S. R.; Sherrill, C. D.; Head-Gordon, M.; Krylov, A . I. J. Chem. Phys., 2000, 113, 3548. Krylov, A . I. J. Chem. Phys., 2000, 113, 6052. Kowalski, K . ; Piecuch, P. J. Chem. Phys., 2000, 113, 18. Kowalski, K . ; Piecuch, P. Chem. Phys. Lett., 2001, 344, 165. Krylov, A . I.; Sherrill, C. D. J. Chem. Phys., 2002, 116, 3194. Dutta, Α.; Sherrill, C. D. J. Chem. Phys., 2003, 118, 1610. Abrams, M. L.; Sherrill, C. D. J. Phys. Chem. A, 2003, 107, 5611. Krylov, A . I. Chem. Phys. Lett., 2001, 338, 375. Krylov, A . I. Chem. Phys. Lett., 2001, 350, 522. Sears, J. S.; Sherrill, C. D.; Krylov, A . J. J. Chem. Phys., 2003, 118, 9084. Sherrill, C. D.; Schaefer, H . F. in Advances in Quantum Chemistry, Löwdin, P.- O., Ed., volume 34, pp 143-269. Academic Press, New York, 1999. Crawford, T. D.; Sherrill, C. D.; Valeev, E. F.; Fermann, J. T.; Leininger, M . L.; King, R. Α.; Brown, S. T.; Janssen, C. L.; Seidl, E. T.; Yamaguchi, Y . ; Allen, W. D.; Xie, Y.; Vacek, C.; Hamilton, P.; Kellogg, C. B . ; Remington, R. B . ; Schaefer III, H . F. PSI 3.0, development version. PSITECH, Inc., Watkinsville, G A 30677, U.S.A., 1999. Abrams, M. L.; Sherrill, C. D. J. Chem. Phys., 2003, 118, 1604. Pople, J. Α.; Binkley, S.; Seeger, R. Int. J. Quantum Chem. Symp., 1976, 10, 1. Purvis, G . D.; Bartlett, R. J. J. Chem. Phys., 1982, 76, 1910. Raghavachari, K . ; Trucks, G . W.; Pople, J. Α.; Head-Gordon, M. Chem. Phys. Lett., 1989, 157, 479. Becke, A . D. J. Chem. Phys., 1993, 98, 1372. Stephens, P. J.; Devlin, F. J.; Chabalowski, C. F.; Frisch, M. J. J. Phys. Chem., 1994, 98, 11623. Roos, B . O.; Taylor, P. R.; Siegbahn, P. Ε. M. Chem. Phys., 1980, 48, 157. Andersson, K . ; Roos, B . O. in Modern Electronic Structure Theory, Yarkony, D . R., Ed., volume 2 of Advanced Series in Physical Chemistry, pp 55-109. World Scientific, Singapore, 1995.


88


28. Olsen, J.; Roos, B . O.; Jørgensen, P.; A a . Jensen, H . J. J. Chem. Phys., 1988, 89, 2185. 29. Grev, R. S.; Schaefer, H. F. J. Chem. Phys., 1992, 96, 6850. 30. Fermann, J. T.; Sherrill, C. D.; Crawford, T. D . ; Schaefer, H . F. J. Chem. Phys., 1994, 100, 8132. 31. Sherrill, C. D.; Schaefer, H . F. J. Phys. Chem., 1996, 100, 6069. 32. Laidig, W. D.; Saxe, P.; Bartlett, R. J. J. Chem. Phys., 1987, 86, 887.


Electron Correlation Methodology - American Chemical Society

Recommend Documents